# What Is the INTEGRITY Stage (2.5) in the ARS Pipeline? Purpose and Implementation

> Discover the INTEGRITY stage 2.5 in the ARS pipeline. Learn how this crucial gate verifies references and citations, preventing fabricated content before peer review.

- Repository: [Edward Cheng-I Wu/academic-research-skills](https://github.com/Imbad0202/academic-research-skills)
- Tags: deep-dive
- Published: 2026-05-13

---

**The INTEGRITY stage (2.5) is the first mandatory gate in the ARS pipeline that verifies references, citations, and data provenance to block fabricated or hallucinated content before any substantive peer review occurs.**

The **INTEGRITY stage (2.5 S)** serves as the primary trust anchor in the Academic Research Skills (ARS) pipeline maintained in the `Imbad0202/academic-research-skills` repository. Located between the planning and review phases, this stage enforces the repository’s "anti-hallucination mandate" by validating scholarly claims through automated verification. According to [`docs/ARCHITECTURE.md`](https://github.com/Imbad0202/academic-research-skills/blob/main/docs/ARCHITECTURE.md) (lines 40-41), no manuscript may proceed to Stage 3 until this gate confirms that citations are reachable, data is consistent, and provenance is documented.

## Core Purpose: Preventing Observer Dilution

The INTEGRITY stage operates as a **hard checkpoint** that prevents "observer dilution"—the risk of reviewers evaluating content without awareness of integrity failures. By requiring explicit user acknowledgement of the Integrity Report before advancement, the pipeline ensures researchers cannot unknowingly propagate AI-generated hallucinations or fabricated citations into peer review.

As implemented in [`docs/ARCHITECTURE.md`](https://github.com/Imbad0202/academic-research-skills/blob/main/docs/ARCHITECTURE.md) (lines 40-41), this stage blocks the pipeline until the user confirms awareness of any detected issues, creating a non-bypassable quality control that maintains scientific credibility.

## How the integrity_verification_agent Validates Claims

The verification logic centers on the **`integrity_verification_agent`**, which executes in distinct modes defined in [`academic-pipeline/references/integrity_review_protocol.md`](https://github.com/Imbad0202/academic-research-skills/blob/main/academic-pipeline/references/integrity_review_protocol.md):

- **Pre-review mode (Stage 2.5)**: Spot-checks approximately **30% of claims** (minimum ten) to verify cited sources are correctly identified, reachable, and match stated content.
- **Final-check mode (Stage 4.5)**: Performs a **zero-tolerance verification** covering 100% of claims before manuscript completion.

The agent evaluates each claim against a **7-mode AI-failure checklist** (defined in [`ai_research_failure_modes.md`](https://github.com/Imbad0202/academic-research-skills/blob/main/ai_research_failure_modes.md)), categorizing specific failure types such as citation fabrication or data misrepresentation.

## The Integrity Report and Material Passport

Upon completion, the agent generates a structured report conforming to **Schema 5** in [`shared/handoff_schemas.md`](https://github.com/Imbad0202/academic-research-skills/blob/main/shared/handoff_schemas.md) (lines 247-249). This document includes:

- **PASS/FAIL outcomes** for each verified claim
- **Suspected AI-failure modes** detected during analysis
- A **Material Passport** recording provenance metadata and verification status

The **Material Passport** creates an auditable trail that follows the manuscript through downstream stages, ensuring transparency in the verification process.

## Integration with the Pipeline Orchestrator

The `integrity_verification_agent` runs immediately after Stage 2 completion, occupying the first transition-time check in the orchestrator’s state machine. According to [`academic-pipeline/agents/pipeline_orchestrator_agent.md`](https://github.com/Imbad0202/academic-research-skills/blob/main/academic-pipeline/agents/pipeline_orchestrator_agent.md) (lines 545-553), the orchestrator dispatches the integrity agent **before** invoking downstream processors like `collaboration_depth_agent`. This sequential enforcement guarantees that synthesis and review agents never process unverified content.

## Zero-Tolerance Final Verification

While Stage 2.5 operates in sampling mode, a second integrity gate at **Stage 4.5 FINAL INTEGRITY** repeats verification at 100% coverage. As noted in [`docs/ARCHITECTURE.md`](https://github.com/Imbad0202/academic-research-skills/blob/main/docs/ARCHITECTURE.md) (line 108), any unresolved failures from Stage 2.5 must be remediated before the final gate can pass, creating a dual-layer validation system.

## Running the INTEGRITY Stage Locally

Execute the integrity gate using the ARS CLI tools:

```bash

# Run pipeline up to the integrity gate (pre-review mode)

ars-full --stage 2.5

# Inspect the generated integrity report

cat ./output/integrity_report.yaml

```

Example report structure:

```yaml
verification_status: PASS
suspect_modes: []
material_passport:
  verification_status: VERIFIED
  repro_lock: null

```

If the report indicates failures, remediate the issues and re-verify:

```bash

# Re-verify after corrections

ars-full --stage 2.5 --reverify

# Proceed to review stage only after user ACK

ars-full --stage 3

```

## Summary

- The **INTEGRITY stage (2.5)** is a non-bypassable gate that verifies scholarly claims before peer review, enforcing the anti-hallucination mandate described in [`README.md`](https://github.com/Imbad0202/academic-research-skills/blob/main/README.md) (lines 536-538).
- The **`integrity_verification_agent`** performs automated spot-checks (30% of claims minimum) against a 7-mode AI-failure checklist during pre-review mode.
- **Schema 5** in [`shared/handoff_schemas.md`](https://github.com/Imbad0202/academic-research-skills/blob/main/shared/handoff_schemas.md) defines the Integrity Report format, including PASS/FAIL outcomes and the Material Passport provenance record.
- **User acknowledgement** is mandatory; the pipeline cannot advance to Stage 3 without explicit confirmation, preventing observer dilution.
- A **zero-tolerance final check** at Stage 4.5 verifies 100% of claims, requiring resolution of all Stage 2.5 failures before manuscript completion.

## Frequently Asked Questions

### What happens if the INTEGRITY stage detects a fabricated citation?

The **`integrity_verification_agent`** flags the claim as FAILED in the Integrity Report, categorizes the specific AI-failure mode from the 7-mode checklist, and blocks progression to Stage 3. You must correct the citation or data issue, then re-run `ars-full --stage 2.5 --reverify` before the pipeline will advance.

### How is the INTEGRITY stage different from the final integrity check at Stage 4.5?

Stage 2.5 operates in **pre-review mode**, sampling approximately 30% of claims (minimum ten) to catch major issues early, while Stage 4.5 FINAL INTEGRITY performs a **zero-tolerance, 100% verification** of all claims before manuscript completion. Any unresolved failures from Stage 2.5 must be fixed before Stage 4.5 can pass, as documented in [`docs/ARCHITECTURE.md`](https://github.com/Imbad0202/academic-research-skills/blob/main/docs/ARCHITECTURE.md) (line 108).

### What is the Material Passport mentioned in the Integrity Report?

The **Material Passport** is a structured metadata record generated during the INTEGRITY stage that documents the provenance of all verified data, citations, and sources. Defined in Schema 5 of [`shared/handoff_schemas.md`](https://github.com/Imbad0202/academic-research-skills/blob/main/shared/handoff_schemas.md) (lines 247-249), it creates an auditable trail of verification status that follows the manuscript through subsequent pipeline stages.

### Can I skip the INTEGRITY stage to speed up the pipeline?

No. The INTEGRITY stage is **non-bypassable by design**. The `pipeline_orchestrator_agent` enforces this as a hard transition-time check (lines 545-553), and the pipeline will not dispatch downstream agents like `collaboration_depth_agent` until the integrity gate passes and the user provides explicit acknowledgement.